Research map / Point clouds / Segmentation and detection
GeoFormer: Geometry Point Encoder for 3D Object Detection with Graph-based Transformer
Xin Jin, Haisheng Su, Cong Ma, Kai Liu, Wei Wu, Fei Hui, Junchi Yan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MLP voxel encoder with GeoFormer (graph-guided transformer) (ablation) | voxel detector (CenterPoint-Pillar) · Waymo val Vehicle L2 (20% train) | 3D AP: 61.02 -> 62.95 | Improved |
| add geometry-guided edge re-weighting to a vanilla transformer encoder (ablation) | GeoFormer · Waymo 20% train | 3D AP (Vehicle/Ped/Cyc): 62.33 -> 62.95; 66.02 -> 67.67; 64.19 -> 65.04 | Improved |
| replace Mamba encoder by GeoFormer (ablation) | GeoFormer · Waymo 20% train | 3D AP Vehicle: 61.90 -> 62.95 | Improved |
| add GPE to PillarNet | PillarNet · nuScenes val | mAP: 55.3 -> 58.9 | Improved |
| add GPE to DSVT | DSVT · nuScenes val | mAP: 66.2 -> 66.5 | Improved |
| add GPE to HEDNet | HEDNet · nuScenes val | mAP: 66.5 -> 67.0 | Improved |
| add GPE to SAFDNet | SAFDNet · nuScenes val | mAP: 66.2 -> 66.7 | Improved |
| add GPE to VoxelMamba | VoxelMamba · nuScenes val | mAP: 67.2 -> 68.2 | Improved |
| use fully connected graph instead of distance-thresholded graph (0.5/2.0) (ablation) | GeoFormer · nuScenes val | mAP / NDS: 66.7 -> 66.4; 71.4 -> 70.8 | Hurt |
| reduce theta_max from 2.0 to 1.5 (ablation) | GeoFormer · nuScenes val | mAP / NDS: 66.7 -> 66.3; 71.4 -> 70.5 | Hurt |
| lower theta_min from 0.8 to 0.5 with theta_max fixed at 2.0 (ablation) | GeoFormer · nuScenes val | mAP / NDS: 66.5 -> 66.7; 71.0 -> 71.4 | Improved |
| replace Dyn/TANet voxel encoder with GPE | CenterPoint-Pillar · nuScenes val | mAP / NDS: 49.9 -> 51.8; 58.5 -> 60.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VoxelMamba — Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection (NeurIPS 2024) · better on 7 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection (CVPR 2018 · similarity 0.90)
- 3D Object Detection With Pointformer (CVPR 2021 · similarity 0.90)
- SEFormer: Structure Embedding Transformer for 3D Object Detection (AAAI 2023 · similarity 0.89)
- MsSVT: Mixed-scale Sparse Voxel Transformer for 3D Object Detection on Point Clouds (NeurIPS 2022 · similarity 0.89)
- Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D Object Detection (AAAI 2024 · similarity 0.89)
- GeoAuxNet: Towards Universal 3D Representation Learning for Multi-sensor Point Clouds (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
